[skyoo2003]

Summary

Derives the gateway's service-alias table from the Smithy models instead of a hand-maintained 63-clause switch, then uses that to onboard the entire AI/Machine Learning category — all 48 upstream models, 42 of them new, with zero hand-written code. Registered services go 105 → 148.

This is Milestone 2 of .claude/prds/aws-service-coverage-100.prd.md ("One category proven end to end"), and it closes that PRD's Open Question 2 (service ID aliasing) and Open Question 3 (is 431 the right denominator).

Related Issue

Refs #136 — this is the follow-on milestone to "repair AWS service onboarding and measure its cost", which measured 3s/service but only for services whose X-Amz-Target prefix already resolved. rekognition failed outright there; it works here.

Fixes #

Changes

Alias derivation (the blocker)

• ir.Model carries the six identifiers a caller can address a service by: shape name, signing name, endpoint prefix, ARN namespace, CloudFormation name, SDK id. Named provider-neutrally — a non-AWS source leaves them empty.
• New codegen.BuildAliases derives alias → service ID and emits internal/generated/aliases/aliases_gen.go. It reports contested aliases rather than resolving them: picking one claimant silently sends the other's traffic to the wrong provider.
• normalizeServiceID is now a 36-entry override map (contested / legacy / substitution, each group documented) plus the generated table. It no longer grows per service.
• selfNamedClaimant settles the common false collision: every runtime split-out signs with its parent's name, so sagemaker is claimed by 8 services — but exactly one of them is sagemaker. That's a fact in the models, not a preference. It resolves 10 of 18 collisions and removed 3 overrides that had been written by hand.

AI/ML category

• 42 new models, generated packages, and scaffolded providers. No hand-written Go.
• 48/48 registered; 17 engine-served or hand-written; 31 registered-only.

Honesty about what "registered" means

• The fidelity manifest now carries each service's Protocol, so registered and covered are separable numbers.
• New docs/coverage.md publishes registered / serving ≥1 / registered-only and never the first alone.
• TestFidelityManifestCoverage used to require every registered service to serve something. That is false by design once rest-json services are registered, so it is now the sharper check: the CRUD engine holds operations for this service and the manifest serves none — which is wiring, and nothing else.

Three defects the category surfaced, each fixed at its root

• smithy.api#Primitive* was unmapped in smithyToGoType, so bedrockagent and omics emitted undefined identifiers and the build failed. All seven prelude primitives are mapped, not only the two that broke.
• make codegen was not idempotent: healthlake.json bundles two namespaces with 8 colliding short names, and model.Shapes is keyed by short name while iterating a Go map. Fixed with sorted iteration plus service-namespace precedence — sorting alone is deterministic and still wrong, because the foreign namespace sorts first.
• One routing regression, caught by the compat suite and missed by unit tests: mapping ses → sesv2 sent every SES v1 Query call to the v2 provider. SES v1 and SESv2 publish identical identifiers and only the protocol separates them, which DetectProtocol already handles.

Test Plan

• CGO_ENABLED=0 go test ./... — all packages pass.
• pytest tests/compatibility — 790 passed, 2 skipped (776 baseline + 14 new).
• make codegen run three times, hashed — byte-identical (eca3ebfd…). This is a new guarantee; it was broken before this PR.
• bash scripts/generate-imports.sh — idempotent.
• All 93 aliases that routed before still route to the same service, pinned by a golden test extracted verbatim from the old switch before it was touched. This is the load-bearing check for replacing hand-written routing with generated routing.
• rekognition end to end: download → codegen → scaffold → build → boto3 ListCollections returns 200, with zero hand-written gateway lines.
• A registered-only service is asserted to answer with an AWS error code, never a fabricated 200 (test_registered_only_service_declines_cleanly).

Measured (105 → 147 services):

	105	147	Extrapolated to 431
Binary	30.8 MiB	31.3 MiB	~35 MiB
Cold start	~460 ms	~460 ms	flat
Idle RSS	57.3 MiB	57.7 MiB	~61 MiB

CI cost at 147: go test ./... 22s, compat suite 9s, codegen 2s, build 1s. The single-binary, zero-config property survives the full 431 target.

Reviewing this PR: 277 files, but 261 are generated packages, downloaded models, or scaffolds. The reviewable surface is 16 files — everything outside internal/generated/, smithy-models/, and internal/services/.

Checklist

• Self-reviewed the code
• Added/updated tests
• Lint/format passes (pre-commit: gofmt, go vet, go build, ruff — all green)
• Updated documentation (docs/coverage.md)
• Added a Changie changelog fragment for user-facing changes

Known gaps

• lex routes nowhere. All four Lex services sign as lex and none is named it; all are rest-json, so none can be engine-served. Each still routes under its own unambiguous name. Distinguishing them needs URL-path routing, which is only worth writing once one of them has a provider.
• amazonrdsv19 and cognito are likewise deliberately unrouted — 2 of 3 and 1 of 2 claimants would be wrong. An empty override value means "decided, and the decision is not to guess".
• The 29 legacy aliases are preserved because deleting them is a silent break for older SDKs, not because a client was observed sending them.

Evidence, with RED/GREEN output verbatim for all 23 guarantees: .claude/tdd/aws-service-coverage-100-milestone-2.tdd.md (local, .claude/ is gitignored).